A data-lean machine learning approach for damage extent estimation and classification in composite structures under multiple failure modes

Abstract

This study addresses the critical challenge of performing accurate structural diagnostics in composite structures under data-scarce and sensor-limited conditions, a known limitation in many existing structural health monitoring (SHM) frameworks. We propose a novel, data-lean simulation-machine learning (ML) approach that integrates high-fidelity numerical modeling with lightweight ML algorithms for simultaneous damage extent estimation and failure mode classification. A time domain spectral finite element model is developed, incorporating physically modeled piezoelectric actuators/sensors and mixed-order layerwise mechanics (both linear and nonlinear), to simulate the electromechanical behavior of composite laminates with multiple concurrent failure mechanisms, namely, fiber breakage, matrix cracking, and delamination. A wide range of representative damage scenarios is constructed, combining intra-laminar and inter-laminar failures distributed through the laminate thickness. These scenarios are simulated using a pitch-and-catch configuration, where a Gaussian pulse is actuated and responses are collected from three spatially distributed sensors. The resulting noisy dataset is used to train ML models for both regression (damage extent) and classification (failure mode), within a unified learning pipeline utilizing interpretable feature representations. The study also explores the impact of data representation strategies and evaluates model performance through cross-validation. Initial results demonstrate strong predictive capabilities, highlighting the potential of simulation-enhanced, ML-enabled SHM systems for robust diagnostics in complex composite structures, even under minimal data and instrumentation constraints.